docs(job-assessment): add architecture, setup and tested prompts

[darthrootbeer]

Context

Last package of the public job assessment system (JBSR-099). The chain already ran offline; what was missing was the docs a stranger needs and prompts that were actually tested, not just written.

Changes

• New ARCHITECTURE.md (every scoring and verdict rule, a worked example, 14 interpretations) and SETUP.md (five-minute install, Claude Code wiring, troubleshooting).
• README rewritten: what it is, what you get, a quickstart verified from a clean clone, the lesson (the model writes findings, scripts check, score and decide, shown with the recorded live-run miss), honest framing and what it does not prove.
• Ten prompts in prompts/, each shown as a block in the docs. tests/test_prompt_blocks.py fails if a block and its file differ, a doc lacks its prompt section or Tested-on line, or a prompt has no saved run and grade.
• Every prompt run on Sonnet and Haiku through the claude CLI. Records and grades in tests/prompt-runs/ (GRADES.md), earlier versions kept. Sonnet passed all ten; Haiku passed eight and partly passed two (Fix errors, and Customize to my background).
• Sanitization: two setup-specific needles removed from a parse_posting test, a hand redaction noted in a live-run record, a wrong relative path fixed in the hooks README. Top-level README and ROADMAP lines that called the docs unfinished are corrected.

Review notes

Private gate (cet-gate.sh) was not available in this session, so it was not run.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XV3Ggh86cf2xwUyz28XaN6

• Docs updated